Feature #18949
Updated by Eregon (Benoit Daloze) over 2 years ago
Ruby has a lot of accidental complexity.
Sometimes it becomes clear some features bring a lot of complexity and yet provide little value or are used very rarely.
Also most Ruby users do not even know about these features.
Replicate and dummy encodings seem to clearly fall into this category, almost nobody uses them but they add a significant complexity and also add a significant performance overhead.
Notably, the existence of those means the number of encodings in a Ruby runtime is actually variable and not fixed.
That means extra synchronization, hashtable lookups, indirections, function calls, etc.
## Replicate Encodings
Replicate encodings are created using `Encoding#replicate(name)`.
It almost sounds like an alias but in fact it is more than that and creates a new Encoding object, which can be used by a String:
```ruby
e = Encoding::US_ASCII.replicate('MY-US-ASCII')
s = "abc".force_encoding(e)
p s.encoding # => e
p s.encoding.name # => 'MY-US-ASCII'
```
This seems completely useless.
There is an obvious first step here which is to change `Encoding#replicate` to return the receiver, and just install an alias for it.
That avoids creating more encoding instances needlessly.
I think we should also deprecate and remove this method though, it is never a good idea to have a global mutable map like this.
If someone want extra aliases for encodings, they can easily to do so by having their own Hash: `{ alias => encoding }.fetch(name) { Encoding.find(name) }`.
## Dummy Encodings
Dummy encodings are not real encodings. They are artificial encodings designed to look like encodings, but don't function as encodings in Ruby.
From the docs:
```
enc.dummy? -> true or false
------------------------------------------------------------------------
Returns true for dummy encodings. A dummy encoding is an encoding for
which character handling is not properly implemented. It is used for
stateful encodings.
```
I wonder why we have those half-implemented encodings in core, it sounds to me like unfinished work which should not have been merged.
The "codepoints" of dummy encodings are just "bytes" and so they behave the same as `Encoding::BINARY`, with the exception of the UTF-16 and UTF-32 dummy encodings.
### UTF-16 and UTF-32 dummy encodings
These two are special dummy encodings.
What they do is they scan the first 2 or 4 bytes of the String, and if those bytes are a byte-order mark (BOM),
the true "actual" encoding is resolved to UTF-16BE/UTF-16LE or UTF-32BE/UTF-32LE.
Otherwise, `Encoding::BINARY` is returned.
This logic is done by `get_actual_encoding()`.
What is weird is this check is not done on String creation, no, it is done *every time* the encoding of that String is accessed (and the result is not stored on the String).
That is a needless overhead and really unreliable semantics.
Do we really want a String which automagically changes between UTF-16LE and UTF-16BE based on mutating its bytes? I think nobody wants that:
```ruby
s = "\xFF\xFEa\x00b\x00c\x00d\x00".force_encoding("UTF-16")
p s # => "\uFEFFabcd"
s.setbyte 0, 254
s.setbyte 1, 255
p s # => "\uFEFF\u6100\u6200\u6300\u6400"
```
I think the path is clear, we should deprecate and then remove Encoding::UTF_16 and Encoding::UTF_32 (dummy encodings).
And then we no longer need `get_actual_encoding()` and the overhead it adds to every String method.
We could also keep those constants and make them refer the native-endian UTF-16/32.
But that could cause confusing errors as we would change the meaning of them.
We could add `Encoding::UTF_16NE` / `Encoding::UTF_16_NATIVE_ENDIAN` if that is useful.
Another possibility would be to resolve these encodings on String creation, like:
```
"\xFF\xFE".force_encoding("UTF-16").encoding # => UTF-16LE
String.new("\xFF\xFE", encoding: Encoding::UTF_16).encoding # => UTF-16LE
"ab".force_encoding("UTF-16").encoding # exception, not a BOM
String.new("ab", encoding: Encoding::UTF_16).encoding # exception, not a BOM
```
I think it is unnecessary to keep such complexity though.
A class method on String or Encoding like e.g. `Encoding.find_from_bom(string)` is so much clearer and efficient (no need to special case those encodings in String.new, #force_encoding, etc).
FWIW JRuby seems to use `getActualEncoding()` only in 2 places (scanForCodeRange, inspect), which is an indication those dummy UTF encodings are barely used if ever. Similarly, TruffleRuby only has 4 usages of `GetActualEncodingNode`.
### Existing dummy encodings
```
> Encoding.list.select(&:dummy?)
[#<Encoding:UTF-16 (dummy)>, #<Encoding:UTF-32 (dummy)>,
#<Encoding:IBM037 (dummy)>, #<Encoding:UTF-7 (dummy)>,
#<Encoding:ISO-2022-JP (dummy)>, #<Encoding:ISO-2022-JP-2 (dummy)>, #<Encoding:ISO-2022-JP-KDDI (dummy)>,
#<Encoding:CP50220 (dummy)>, #<Encoding:CP50221 (dummy)>]
```
So besides UTF-16/UTF-32 dummy, it's only 7 encodings.
Does anyone use one of these 7 dummy encodings?
What is interesting to note, is that these encodings are exactly the ones that are also not ASCII-compatible, with the exception of UTF-16BE/UTF-16LE/UTF-32BE/UTF-32LE (non-dummy).
As a note, UTF-{16,32}{BE,LE} are ASCII-compatible in codepoints but not in bytes, and Ruby uses the bytes definition of ASCII-compatible.
There is potential to simplify encoding compatibility rules and encoding compatibility checks based on that.
So what this means is if we removed dummy encodings, all encodings except UTF-{16,32}{BE,LE} would be ASCII-compatible, which would lead to significant simplifications for many string operations which currently need to handle dummy encodings specially.
Unicode encodings like UTF-{16,32}{BE,LE} already have special behavior for some Ruby methods, so those are already handled specially in some places (they are the only encodings with minLength > 1).
```
> Encoding.list.reject(&:ascii_compatible?)
[#<Encoding:UTF-16BE>, #<Encoding:UTF-16LE>,
#<Encoding:UTF-32BE>, #<Encoding:UTF-32LE>,
#<Encoding:UTF-16 (dummy)>, #<Encoding:UTF-32 (dummy)>,
#<Encoding:IBM037 (dummy)>, #<Encoding:UTF-7 (dummy)>,
#<Encoding:ISO-2022-JP (dummy)>, #<Encoding:ISO-2022-JP-2 (dummy)>, #<Encoding:ISO-2022-JP-KDDI (dummy)>,
#<Encoding:CP50220 (dummy)>, #<Encoding:CP50221 (dummy)>]
```
What can we do with such a dummy non-ASCII-compatible encoding?
Almost nothing useful:
```ruby
s = "abc".encode("IBM037")
=> "\x81\x82\x83"
> s.bytes
=> [129, 130, 131]
> s.codepoints
=> [129, 130, 131]
> s == "abc"
=> false
> "été".encode("IBM037")
=> "\x51\xA3\x51"
```
So about the only thing that works with them is `String#encode`.
I think we could preserve that functionality, if actually used (does anyone use one of these 7 dummy encodings?), through:
```ruby
> "été".encode("IBM037")
=> "\x51\xA3\x51" (.encoding == BINARY)
> "\x51\xA3\x51".encode("UTF-8", "IBM037") # encode from IBM037 to UTF-8
=> "été" (.encoding == UTF-8)
```
That way there is no need for those to be Encoding instances, we would only need the conversion tables.
It is even better if we can remove them, so the notion of "dummy encodings" can disappear completely and nobody needs to understand or implement them.
### rb_define_dummy_encoding(name)
The C-API has `rb_define_dummy_encoding(const char *name)`.
This creates a new Encoding instance with `dummy?=true`, and it is also non-ASCII-compatible.
There seems to be no purpose to this besides storing the metadata of an encoding which does not exist in Ruby.
This seems a really expensive/complex way to handle that from the VM point of view (because it dynamically creates an Encoding and add it to lists/maps/etc).
A simple replacement would be to mark the String as BINARY and save the encoding name as an instance variable of that String.
Since anyway Ruby can't understand anything about that String, it's just raw bytes to Ruby's eyes.
## Summary
I suggest we deprecate replicate and dummy encodings in Ruby 3.2.
And then we remove them in the next version.
This will significantly simplify string-related methods, and the behavior exposed to Ruby users.
It will also significantly speedup encoding lookup in CRuby (and other Ruby implementations).
With a fixed number of encodings we can ensure all encoding indices fit in 7 bits, and `ENCODING_GET` can be simply `RB_ENCODING_GET_INLINED`.
`get_actual_encoding()` will be gone and its overhead as well.
`rb_enc_from_index()` would be just `return global_enc_table->list[index].enc;`, instead of the expensive behavior currently with `GLOBAL_ENC_TABLE_EVAL` which takes a lock and more when there are multiple Ractors.
Many checks in these methods would be removed as well.
Yet another improvement would be to load all encodings eagerly, that is small and fast in my experience, what is slow and big is the conversion tables, that'd simplify `must_encindex()` further.
These changes would affect most String methods, which use
```
STR_ENC_GET->get_encoding which does:
get_actual_encoding->rb_enc_from_index and possibly ->enc_from_index
ENCODING_GET->RB_ENCODING_GET_INLINED and possibly ->rb_enc_get_index->enc_get_index_str->rb_attr_get
```
Some of these details are mentioned in https://github.com/ruby/ruby/pull/6095#discussion_r915149708.
The overhead is so large that it is worth handling some hardcoded encoding indices directly in String methods.
This feels wrong, getting the encoding from a String should be simple, straightforward and fast.
Further optimizations will be unlocked as the encoding list becomes fixed and immutable.
For example, the name-to-Encoding map is then immutable and could use perfect hashing.
Inline caching those lookups also becomes easier as the the map cannot change.
Also that map would no longer need synchronization, etc.
## To Decide
Each item is independent. I think 1 & 2 are very important, 3 less but would be nice.
1. Remove `Encoding#replicate` and `rb_define_dummy_encoding()`. With that there is a fixed number of encodings, a lot of simplifications and many optimizations become available. They are used respectively in only 1 gem and 5 gems, see https://bugs.ruby-lang.org/issues/18949#note-4
2. Remove the dummy UTF-16 and UTF-32 encodings. This removes the need for `get_actual_encoding()` which is expensive. This functionality seems rarely used in practice, and it only works when such strings have a BOM, which is very rare.
3. Remove other dummy encodings, so there are no more dummy "half-implemented" encodings and all encodings are ASCII-compatible in terms of codepoints.